trtx-rs

⚠️ EXPERIMENTAL - NOT FOR PRODUCTION USE

This project is in early experimental development. The API is unstable and will change. This is NOT production-ready software. Use at your own risk.

Published on crates.io to reserve the crate names.

Safe Rust bindings to NVIDIA TensorRT-RTX for high-performance deep learning inference.

Overview

This project provides ergonomic Rust bindings to TensorRT-RTX, enabling efficient inference of deep learning models on NVIDIA GPUs with minimal overhead.

Features

Safe API: RAII-based memory management and type-safe abstractions
Two-phase workflow: Separate build (AOT) and inference (runtime) phases
Zero-cost abstractions: Minimal overhead over C++ API
Comprehensive error handling: Proper Rust error types for all operations
Flexible logging: Customizable log handlers for TensorRT messages

Project Structure

trtx-rs/
├── trtx-sys/       # Raw FFI bindings (unsafe)
└── trtx/           # Safe Rust wrapper (use this!)

Prerequisites

Required (building)

NVIDIA TensorRT-RTX 1.3: Download and install from NVIDIA Developer
CUDA Runtime: Version compatible with your TensorRT-RTX installation
Clang: Required for autocxx. On Windows: winget install LLVM.LLVM
NVIDIA GPU: Compatible with TensorRT-RTX requirements

TensorRT is by default dynamically loaded. So, the TensorRT SDK is only required for building with Cargo features link_tensorrt_rtx/ link_tensorrt_onnxparser which would link the TensorRT libraries. Use TENSORRT_RTX_DIR to point to the TensorRT SDK root directory (the path that contains the lib folder with the shared libraries).

Required (GPU execution)

NVIDIA TensorRT-RTX: Download and install from NVIDIA Developer
- The TensorRT libraries should be in a location where they can be dynamically loaded. (e.g. by setting PATH on Windows or LD_LIBRARY_PATH on Linux)
- This crate currently requires TensorRT RTX version 1.3 (see Cargo feature v_1_3). Other versions, might become available in future.
NVIDIA GPU: Compatible with TensorRT-RTX requirements

Development Without TensorRT-RTX (Mock Mode)

If you're developing on a machine without TensorRT-RTX (e.g., macOS, or for testing), you can use the mock feature. This enables the trtx mock layer (safe Rust stubs in trtx that mirror the real API), not the low-level FFI:

# Build with mock mode
cargo build --features mock

# Run examples with mock mode
cargo run --features mock --example basic_workflow

# Run tests with mock mode
cargo test --features mock

Mock mode provides stub implementations that allow you to:

Verify the API compiles correctly
Test your application structure
Develop without needing an NVIDIA GPU
Run CI/CD pipelines on any platform

Note: Mock mode only validates structure and API usage. For actual inference, you need real TensorRT-RTX.

Cargo features

The trtx crate has the following Cargo features:

default: "real", "dlopen_tensorrt_onnxparser", "dlopen_tensorrt_rtx", "onnxparser", "v_1_3"
mock: use this library in mock mode. TensorRT libraries and a Nvidia are no longer necessary for execution
real: opposite of mock mode. TensorRT and Nvidia GPU are required for execution
dlopen_tensorrt_rtx: enables dynamic loading of the TensorRT library via trtx::dynamically_load_tensorrt
dlopen_tensorrt_onnxparser: enables dynamic loading of the TensorRT ONNX parser library via trtx::dynamically_load_tensorrt_onnxparser
links_tensorrt_rtx: links the TensorRT library, trtx::dynamically_load_tensorrt is now optional
links_tensorrt_onnxparser: links the TensorRT ONNX parser library, trtx::dynamically_load_tensorrt_onnxparser is now optional
onnxparser: Enables the ONNX parser functionality of this crate. Optional if not using ONNX as the input format for TensorRT, but using the builder library instead
v_1_3: Needs to be always enabled. Future TensorRT versions might be selectable by higher version numbers in future

Installation

Add to your Cargo.toml:

[dependencies]
trtx = "0.3"

Usage

Build Phase (Creating an Engine)

use trtx::{Logger, Builder};
use trtx::builder::{network_flags, MemoryPoolType};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Dynamically load TensorRT with optional path
    // when using the crate's dlopen_tensorrt_rtx feature (the default, no-op when link_tensorrt_rtx is also enabled)
    trtx::dynamically_load_tensorrt(None::<String>).unwrap();

    // Create logger
    let logger = Logger::stderr()?;

    // Create builder
    let builder = Builder::new(&logger)?;

    // Create network with explicit batch dimensions
    let network = builder.create_network(network_flags::EXPLICIT_BATCH)?;

    // Configure builder
    let mut config = builder.create_config()?;
    config.set_memory_pool_limit(MemoryPoolType::Workspace, 1 << 30)?; // 1GB

    // Build serialized engine
    let engine_data = builder.build_serialized_network(&network, &config)?;

    // Save to disk
    std::fs::write("model.engine", &engine_data)?;

    Ok(())
}

Inference Phase (Running Inference)

use trtx::{Logger, Runtime};
use std::fs;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Dynamically load TensorRT with optional path
    // when using the crate's dlopen_tensorrt_rtx feature (the default, no-op when link_tensorrt_rtx is also enabled)
    trtx::dynamically_load_tensorrt(None::<String>).unwrap();

    // Create logger and runtime
    let logger = Logger::stderr()?;
    let runtime = Runtime::new(&logger)?;

    // Load serialized engine
    let engine_data = fs::read("model.engine")?;
    let engine = runtime.deserialize_cuda_engine(&engine_data)?;

    // Create execution context
    let mut context = engine.create_execution_context()?;

    // Query tensor information
    let num_tensors = engine.get_nb_io_tensors()?;
    for i in 0..num_tensors {
        let name = engine.get_tensor_name(i)?;
        println!("Tensor {}: {}", i, name);
    }

    // Set tensor addresses (requires CUDA memory)
    unsafe {
        context.set_tensor_address("input", input_device_ptr)?;
        context.set_tensor_address("output", output_device_ptr)?;
    }

    // Execute inference
    unsafe {
        context.enqueue_v3(cuda_stream)?;
    }

    Ok(())
}

Custom Logging

use trtx::{Logger, LogHandler, Severity};

struct MyLogger;

impl LogHandler for MyLogger {
    fn log(&self, severity: Severity, message: &str) {
        match severity {
            Severity::Error | Severity::InternalError => {
                eprintln!("ERROR: {}", message);
            }
            Severity::Warning => {
                println!("WARN: {}", message);
            }
            _ => {
                println!("INFO: {}", message);
            }
        }
    }
}

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let logger = Logger::new(MyLogger)?;
    // Use logger...
    Ok(())
}

API Overview

Core Types

Logger: Captures TensorRT messages with custom handlers
Builder: Creates optimized inference engines
NetworkDefinition: Defines the computational graph
BuilderConfig: Configures optimization parameters
Runtime: Deserializes engines for inference
CudaEngine: Optimized inference engine
ExecutionContext: Manages inference execution

Error Handling

All fallible operations return Result<T, Error>:

use trtx::Error;

match builder.create_network(0) {
    Ok(network) => {
        // Use network
    }
    Err(Error::InvalidArgument(msg)) => {
        eprintln!("Invalid argument: {}", msg);
    }
    Err(e) => {
        eprintln!("Error: {}", e);
    }
}

Safety

Safe Operations

Most operations are safe and use RAII for resource management:

Creating loggers, builders, runtimes
Building and serializing engines
Deserializing engines
Creating execution contexts

Unsafe Operations

CUDA-related operations require unsafe:

set_tensor_address: Must point to valid CUDA device memory
enqueue_v3: Requires valid CUDA stream and properly bound tensors

Building from Source

# Clone the repository
git clone https://github.com/rustnn/trtx-rs.git
cd trtx-rs

# Option 1: Build with TensorRT-RTX (requires NVIDIA GPU)
export TENSORRT_RTX_DIR=/path/to/tensorrt-rtx
cargo build --release
cargo test

# Option 2: Build in mock mode (no GPU required)
cargo build --features mock --release
cargo test --features mock
cargo run --features mock --example basic_workflow

Examples

See the trtx/examples/ directory for complete examples:

basic_workflow.rs: Build and serialize an engine (optionally from ONNX), then run inference
tiny_network.rs: Build a small ReLU-based network from scratch using the Network API (no ONNX)
rustnn_executor.rs: rustnn-compatible executor integration

Architecture

trtx-sys (FFI Layer)

autocxx-generated bindings for the TensorRT-RTX C++ API
Slim C++ logger bridge for virtual method handling (e.g., log callbacks)
Optional mock FFI (when mock feature is enabled) so the crate can build without TensorRT installed
No safety guarantees; internal use only

trtx (Safe Wrapper)

Mock layer: When the mock feature is enabled, the trtx crate uses a Rust mock layer (trtx/src/mock/) that mirrors the real API—this is the “mock mode” you use for development without GPU. Real implementation lives in trtx/src/real/.
RAII-based resource management
Type-safe API
Lifetime tracking
Comprehensive error handling
User-facing API

Troubleshooting

Build Errors

Cannot find TensorRT headers:

export TENSORRT_RTX_DIR=/path/to/tensorrt-rtx

Linking errors:

export LD_LIBRARY_PATH=$TENSORRT_RTX_DIR/lib:$LD_LIBRARY_PATH

Runtime Errors

CUDA not initialized: Ensure CUDA runtime is properly initialized before creating engines or contexts.

Invalid tensor addresses: Verify that all tensor addresses point to valid CUDA device memory with correct sizes.

Development

Pre-commit Hooks

To ensure code quality, set up the pre-commit hook:

cp .githooks/pre-commit .git/hooks/pre-commit
chmod +x .git/hooks/pre-commit

The hook will automatically run cargo fmt and cargo clippy before each commit.

Manual Checks

You can also run checks manually using the Makefile:

make check-all  # Run fmt, clippy, and tests
make fmt        # Format code
make clippy     # Run lints
make test       # Run tests

GPU Testing

The project includes CI workflows for testing with real NVIDIA GPUs:

Mock mode CI: Runs on every push (ubuntu, macos) - tests API without GPU
GPU tests: Runs on self-hosted Windows runner with T4 GPU - tests real TensorRT-RTX

To set up a GPU runner for real hardware testing, see GPU Runner Setup Guide.

The GPU tests workflow:

Builds without mock mode (uses real TensorRT-RTX)
Verifies CUDA and GPU availability
Runs tests and examples with actual GPU acceleration
Can be triggered manually or runs automatically on code changes

Contributing

Contributions are welcome! Please see docs/DESIGN.md for architecture details.

License

This project is licensed under the Apache License, Version 2.0 - see the LICENSE file for details.

Acknowledgments

NVIDIA for TensorRT-RTX
The Rust community for excellent FFI tools

Status

This project is in early development. APIs may change before 1.0 release.

Implemented

✅ Core FFI layer (autocxx); trtx mock layer for development without TensorRT (no GPU)
✅ Logger interface with custom handlers
✅ Builder API for engine creation
✅ Runtime and engine deserialization
✅ Execution context
✅ Error handling with detailed messages
✅ Network API: TensorRT-RTX INetworkDefinition supported—build networks in Rust without ONNX
✅ ONNX parser bindings (nvonnxparser integration)
✅ CUDA: cudarc integration for memory management and device sync
✅ rustnn-compatible executor API (ready for integration)
✅ RAII-based resource management

Planned

⬜ Dynamic shape support
⬜ Optimization profiles
⬜ Weight refitting
⬜ INT8 quantization support
⬜ Comprehensive examples with real models
⬜ Performance benchmarking
⬜ Documentation improvements

trtx-sys 0.3.0